[sernote]

The Anthropic checklist attributes thinking-block stripping after an ordinary user message to every Haiku model. Current documentation preserves Haiku 5.5 thinking blocks and limits the older rule to Haiku through 4.5. That categorical instruction can give a multi-turn cache audit the wrong starting cause.

Correct the two model boundaries in the existing checklist line. The inclusive older-model bound is written as Haiku ≤4.5; the surrounding guidance is preserved. Sources: prompt caching, October 7 release notes.

Validation:

• 260 unit tests pass; package validator, 40 trigger cases, Python syntax, whitespace and bytecode checks pass.
• Deferred corpus: 381,258 → 381,260 characters; aggregate estimate stays 95,315 under the unchanged test ceiling. Static plugin-eval score remains 59; trigger/invoke/deferred estimates and failure set are unchanged.
• Two frozen advisory cases (Haiku 5.5 and 4.5), once before and once after, were correct in both phases. After sessions used the equivalent long-form wording; separate final review verified the compact notation's equivalence, clarity, scope and budget. Independent product review: PASS.

The benefit demonstrated here is literal source alignment. No behavioral uplift or production cache benefit was measured. The diff contains one reference line.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-08T15:29:23.900058Z	607ea76	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.